Model Description
- Base model:
Qwen/Qwen3-8B
- Format: Hugging Face Transformers + Safetensors
- Intended domain: web navigation, browser interaction, reasoning-to-action generation
Intended Uses
This model is intended for:
- Research on language agents for web environments
- Multi-turn prompting with reasoning and action generation
- Offline evaluation on browser-task datasets
- Experiments involving tool-use and UI-grounded planning
Out-of-Scope Uses
This model is not intended for:
- Production autonomy without action validation
- Safety-critical or regulated domains
- Situations requiring guaranteed factuality or correctness
- Fully trusted execution of generated actions
Training Data
This model was fine-tuned on a WEASEL/AgentTrek-style web-agent dataset with message-based interaction trajectories.
The underlying trajectories come from yeonjooooni/agenttrek-WEASEL, while the reasoning traces were newly generated using inference from Qwen3-8B.
Training Objective
The fine-tuning objective emphasizes:
- following browser-task instructions
- producing coherent next actions
- grounding outputs in observed state
- supporting reasoning-heavy interaction traces
Limitations
- The model can generate invalid or non-executable actions
- Reasoning traces may sound confident while being wrong
- Performance may degrade outside the web-agent data distribution
- Prompt wording can materially affect output quality
- Additional runtime checks are recommended before action execution
Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "yeonjooooni/Qwen3_8B_WEASEL"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
